KR20220023545A - Apparatus and method for diagnosing partial discharge - Google Patents

Abstract

Embodiments of the present disclosure relate to an apparatus and method for diagnosing partial discharge of a power facility based on a self-attention mechanism. An electronic apparatus includes a memory storing a learning model and an analysis unit configured to diagnose partial discharge by using the learning model. The analysis unit may be configured to generate embedding data based on a pre-defined partial discharge signal pattern (Phase-resolved PD), extract a characteristic from the embedding data based on the self-attention mechanism, and diagnose partial discharge based on a result of the extraction.

Description

Partial discharge diagnostic method and device

Various embodiments of the present disclosure relate to a method and apparatus for diagnosing partial discharge of power equipment, and more particularly, to a method and apparatus for diagnosing partial discharge based on a self-attention mechanism.

Power facilities such as power plants and substations include a variety of power devices such as a generator, a gas insulated switchgear, a gas insulated transformer, and an oil immersed transformer.

Since a high current flows in such a power device, partial discharge occurs as a precursor to a failure of the power device. Partial discharge is an incomplete dielectric breakdown phenomenon that does not completely bridge between electrodes in an electric device. noise), etc.

Recently, partial discharge diagnosis methods using various algorithms have been proposed in order to minimize damage to social property and human life from power equipment accidents. Representative models used in the partial discharge diagnosis method include an artificial neural network, a convolutional neural network, a recurrent neural network (RNN), and a long short term memory (LSTM).

However, such an algorithm requires many parameter settings such as momentum coefficient, number of repetitions, learning rate, and number of nodes, performance may vary greatly depending on parameter changes, and has a limit in the speed of diagnosing partial discharge.

The present disclosure has been made to solve the above problems, and an object of the present disclosure is to provide a method and an apparatus for diagnosing partial discharge based on a self-attention mechanism.

The technical problems to be achieved in the present disclosure are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by those of ordinary skill in the art to which the present disclosure belongs from the description below. There will be.

A method for diagnosing partial discharge according to various embodiments of the present disclosure is characterized by generating embedding data based on a predefined partial discharge signal pattern (Phase-resolved PD) and from the embedding data based on a self-focusing technique It may include an operation of extracting , and an operation of diagnosing partial discharge based on the extraction result.

An electronic device according to various embodiments of the present disclosure includes a memory for storing a learning model and an analyzer configured to diagnose partial discharge using the learning model, wherein the analyzer includes a predefined partial discharge signal pattern (Phase-resolved). PD), extracting features from the embedding data based on a self-focusing technique, and diagnosing a partial discharge based on a result of the extraction.

An apparatus and method for diagnosing partial discharge according to embodiments of the present disclosure check a correlation between input data and output data based on a self-concentration technique and use this for learning to diagnose partial discharge, thereby improving diagnostic performance of partial discharge and improving partial discharge It is possible to improve the complexity of the discharge diagnosis algorithm.

Effects obtainable in the present disclosure are not limited to the above-mentioned effects, and other effects not mentioned may be clearly understood by those of ordinary skill in the art to which the present disclosure belongs from the description below. will be.

1 is a block diagram of an electronic device for diagnosing partial discharge according to various embodiments of the present disclosure;
2A to 2C are diagrams conceptually illustrating an analysis unit according to various embodiments of the present disclosure.
3 is a flowchart illustrating a partial discharge diagnosis operation of an electronic device according to various embodiments of the present disclosure;
4 is a flowchart illustrating an operation of extracting a feature of embedding data in an electronic device according to various embodiments of the present disclosure.
5 is a flowchart illustrating an operation of diagnosing partial discharge in an electronic device according to various embodiments of the present disclosure;
6A and 6B are diagrams for explaining the performance of partial discharge diagnosis according to the present disclosure.

Hereinafter, various embodiments will be described in detail with reference to the accompanying drawings. Regardless of the reference numerals, the same or similar components may be assigned the same reference numbers and duplicate descriptions thereof may be omitted.

The suffix 'module' or 'part' for the components used in the following description is given or used interchangeably for ease of writing the specification, and does not have a meaning or role distinct from each other by itself. In addition, 'module' or 'unit' refers to software or hardware components such as field programmable gate array (FPGA) or application specific integrated circuit (ASIC), but is not limited to software or hardware. A 'unit' or 'module' may be configured to reside on an addressable storage medium or may be configured to reproduce one or more processors. Thus, as an example, 'part' or 'module' refers to components such as software components, object-oriented software components, class components and task components, processes, functions, properties, may include procedures, subroutines, segments of program code, drivers, firmware, microcode, circuitry, data, databases, data structures, tables, arrays, and variables. The functionality provided in one component, 'unit' or 'module' may be combined into a smaller number of components and 'unit' or 'module' or additional components and 'unit' or 'module' can be further separated.

The steps of a method or algorithm described in connection with some embodiments of the present disclosure may be implemented directly in hardware executed by a processor, a software module, or a combination of the two. A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of recording medium known in the art. An exemplary recording medium is coupled to the processor, the processor capable of reading information from, and writing information to, the storage medium. Alternatively, the recording medium may be integral with the processor. The processor and recording medium may reside within an application specific integrated circuit (ASIC). The ASIC may reside within the user terminal.

Terms including an ordinal number such as 1st, 2nd, etc. may be used to describe various elements, but the elements are not limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

When a component is referred to as being 'connected' or 'connected' to another component, it is understood that it may be directly connected or connected to the other component, but other components may exist in between. it should be On the other hand, when it is mentioned that a certain element is 'directly connected' or 'directly connected' to another element, it should be understood that the other element does not exist in the middle.

First, the terms used in this specification will be briefly described.

Artificial intelligence is a computer system or device equipped with human intelligence, and may refer to artificially implemented human intelligence in a machine or the like. Artificial intelligence also refers to the field of science that studies the methodology or feasibility of creating intelligence.

Artificial neural network is a model or algorithm that implements artificial intelligence. It is a statistical learning algorithm modeled by simulating the neural network of biology in machine learning. Therefore, it can be called a model or learning algorithm with problem-solving ability.

An artificial neural network may include an input layer, an output layer, and one or more hidden layers. Each layer of the artificial neural network includes a plurality of nodes corresponding to neurons of the neural network, and a node in one layer of the artificial neural network and a node in another layer may be connected by a synapse. As an embodiment, an artificial neural network in which all nodes of each layer and all nodes of the next layer are synaptically connected may be referred to as a fully connected artificial neural network.

In the artificial neural network, each node may receive input signals input through a synapse and generate an output value based on an activation function for weights and biases for each input signal.

A deep neural network may collectively refer to an artificial neural network including a plurality of hidden layers between an input layer and an output layer. A deep neural network can model complex nonlinear relationships and can have various structures depending on its purpose. For example, as a deep neural network structure, there may be a convolutional neural network, a recurrent neural network (RNN), a long short term memory (LSTM), and the like.

The convolutional neural network may be effective in classifying and identifying images and/or moving images by learning by identifying features of structured spatial data such as images, moving images, and character strings. The recurrent neural network has a cyclic structure inside, so the learning of the past time is multiplied by the weight and reflected in the current learning. perform the function Therefore, it can be effective in learning sequential data to perform classification or prediction.

LSTM is a type of recurrent neural network, which solves the problem that old data of the cyclical neural network disappears without affecting it. Like the cyclical neural network, it can be effective in performing classification or prediction by learning sequential data.

1 is a block diagram of an electronic device 100 for diagnosing partial discharge according to various embodiments of the present disclosure.

Referring to FIG. 1 , the electronic device 100 may include a processor 110 , a memory 120 , a sensing unit 130 , an analysis unit 140 , and an output unit 150 . The electronic device 100 may include, for example, at least one of a robot device, a portable communication device, a computer device, a portable multimedia device, a portable medical device, a camera, a wearable device, and a home appliance device.

Each of the components of the electronic device 100 illustrated in FIG. 1 may be composed of one chip, component, or electronic circuit, or a combination of chips, components, or electronic circuit. According to another embodiment, some of the components shown in FIG. 1 may be divided into a plurality of components (eg, a plurality of processors) and configured as different chips or parts or electronic circuits, and some components They may be combined to form a single chip, component or electronic circuit. In addition, at least one of the above-described components may be omitted or other components may be added to the electronic device 100 .

According to various embodiments, the processor 110 may perform data processing and/or calculation for the overall operation of the electronic device 100 . The processor 110 may control at least one other component included in the electronic device 100 by driving a software program. In addition, the processor 110 may perform learning according to machine learning according to the program code stored in the memory 120 , and may control to store or output the learning result in the memory 120 .

According to various embodiments, the memory 120 includes at least one component (eg, the processor 110 , the memory 120 , the sensing unit 130 , the analysis unit 140 , or the output unit of the electronic device 100 ). (150)) can store various data used by. The data may include, for example, input data or output data for software (eg, a program) and instructions related thereto. The memory 120 may include a volatile memory or a non-volatile memory.

According to various embodiments, the sensing unit 130 may detect a partial discharge (PD) signal of a power facility (eg, a power device). According to an embodiment, the sensing unit 130 may include a High Frequency Current Transformer (HFCT), a Radio Frequency Current Transformer (RFCT) sensor, etc. However, this is only an example, and the present disclosure is not limited thereto. For example, various sensors or various devices capable of detecting a partial discharge signal of a power facility may be used as the configuration of the sensing unit 130 .

According to various embodiments, the analysis unit 140 may diagnose the partial discharge by extracting features from the partial discharge signal detected by the sensing unit 130 . Diagnosing the partial discharge may include checking whether the partial discharge has occurred, a type of the partial discharge, and the like. The type of partial discharge may include at least one of particles, floating, corona, vibration, and noise. According to an embodiment, the analysis unit 140 infers a result from the input data X using a learning model, and classifies the partial discharge based on the inferred result, as will be described later with reference to FIGS. 2A to 2C . can do. In this case, when inferring a result, the analyzer 140 may use a self-attention mechanism that calculates a correlation between the input data and the output data by referring to the input data once again.

According to various embodiments, the output unit 150 may generate an output related to the operation of the electronic device 100 . The output may include partial discharge diagnostic results. According to an embodiment, the output unit 150 may include a display that outputs visual information, an audio data output device that outputs auditory information, and a haptic module that outputs tactile information. For example, the display may include a liquid crystal display (LCD), a light emitting diode (LED) display, an organic light emitting diode (OLED) display, or a microelectromechanical system (MEMS) display, or electronic paper or the like. . Also, the audio data output device may include at least one of a speaker, an earphone, an earphone, and a headset included in the electronic device 100 or connected to the electronic device 100 through wired/wireless. Additionally or alternatively, the display may include touch circuitry configured to sense a touch, or sensor circuitry configured to measure the intensity of a force generated by the touch (eg, a pressure sensor).

According to various embodiments, the analysis unit 140 may be implemented as hardware, software, or a combination of hardware and software. In the above-described embodiment, it has been described that the analysis unit 140 is implemented separately from the processor 110 and the memory 120, but the present disclosure is not limited thereto. For example, the analysis unit 140 may be implemented by being integrated with the processor 110 or integrated with the memory 120 .

2A to 2C are diagrams conceptually illustrating the analysis unit 140 according to various embodiments of the present disclosure.

Referring to FIG. 2A , the analysis unit 140 may include a self-attention block 200 and a classification block 210 . The self-concentration block 200 may output F by inferring a result for the input data X based on the learning model, and the classification block 210 is a partial discharge based on the output of the self-concentration block 200 . can be diagnosed In this case, as described above, the self-focusing block 200 may infer a result using the self-focusing technique and use it for fault diagnosis.

According to various embodiments, the self-concentration block 200 configured to infer and output the result for the input data X, as shown in FIG. 2B , includes a plurality of self-focusing sub-blocks (multi-head self). -attention sub-block) 202 and may be composed of a feed forward block (Feed Forward block) (206). However, this is merely exemplary, and the present disclosure is not limited thereto. For example, the implementation of the self-concentration block 200 with a plurality of self-concentration sub-blocks 202 is to extract various features of the embedding data, and the self-concentration block 200 is a single self-concentration sub-block 202 . It may be configured, and a plurality of self-concentration blocks 200 may be implemented.

According to various embodiments, the self-concentration block 200 may generate N-dimensional embedding data X in order to infer a result. The embedding data X may be a vector having N real values, and may include a query vector, a key vector, and a value vector. For example, the self-focusing block 200 may generate a key vector, a value vector, and a query vector based on a partial discharge signal pattern (Phase-resolved PD) obtained from a simulated failure test. According to an embodiment, the embedding data X may be generated based on an algorithm such as one-hot encoding, word2vec, or GloVe.

According to various embodiments, when the embedding data X is generated, the self-concentration block 200 uses a plurality of self-concentration sub-blocks 202 to calculate an attention value for the embedding data X. Embedding data can be learned. For example, the self-concentration block 200 may calculate the attention value in parallel by dividing the received embedding data X by the number of the self-concentration sub-blocks 202 . The attention value may indicate a relationship between a query vector and a key vector.

According to an embodiment, the embedding data X may be composed of a query vector and a key vector having a dk dimension, and a value vector having a dv dimension. Using each of the sub-blocks 202, as shown in Equation 1 below, a similarity obtained by dot-product of a query vector and a key vector is scaled (scaled dot-product), and this is a softmax function After normalization by applying to , the attention value can be calculated by multiplying it with a value vector.

Referring to <Equation 1>, T denotes a transpose matrix, and a value vector similar to a query vector may have a higher attention value.

According to an embodiment, the self-concentration block 200 merges the attention values calculated by using each self-concentration sub-block 202 as shown in Equation 2 below, so that the embedding data X is It can be calculated as an attention value.

According to various embodiments, when the attention value is calculated, the self-concentration block 200 may output the calculated attention value to the classification block 210 through the feed forward block 206 . For example, the self-concentration block 200 may apply ( 204 ) a residual connection to the attention value calculated by the self-concentration sub-block 202 , and output the result to the feed forward block 206 . Residual concatenation is to prevent the learning data learned in the self-concentration sub-block 202 from being lost in the data processing process. Residual concatenation that adds the output data of (202) and the input data (eg, embedding data (X)) can be applied.

In this case, the feed forward block 206 may linearly transform the calculated attention value to output the calculated attention value to the classification block 210 . For example, the feed forward block 206 linearly transforms the attention value to which the residual connection is applied using a nonlinear function (ReLU), and applies the residual connection 208 to the linearly transformed attention value to the classification block 210 . can be printed out. The output F of the feed forward block 206 may be defined as in Equation 4 below.

In summary, the self-concentration block 200 takes a partial discharge signal as an input, extracts a correlation between the input and the output, and uses it for learning. It may be defined as follows, and an output when configured to learn using a plurality of self-concentration blocks 200 may be defined as in Equation 6 below.

According to various embodiments, the classification block 210 configured to diagnose the partial discharge based on the output of the self-concentration block 200 may include a pooling layer 212 and a dense layer, as shown in FIG. 2C . (dense layer) 214 and may be composed of a softmax layer (softmax layer) (216).

According to various embodiments, the data output by the self-concentration block 200 may be input to the classification block 210 , and the classification block 210 may adjust the size of the input data using the polling layer 212 . there is. In other words, the classification block 210 may adjust the size of the data in order to prevent overfitting of the input data and reduce the number of parameters by using the pooling layer 212 .

According to an embodiment, the classification block 210 divides input data into filters of a predetermined size and selects the largest value in the filter as shown in Equation 7 below, based on a max pooling technique Thus, the size of the input data can be reduced. However, this is only an example, and the present disclosure is not limited thereto. For example, the classification block 210 may use various types of pooling techniques capable of preventing overfitting of input data and reducing the number of parameters, such as an average pooling technique of selecting an average value in a filter.

According to various embodiments, when the size of the input data is adjusted, the classification block 210 may generate a fault representation vector using the density layer 214 . In this case, the classification block 210 may generate a defect expression vector using a non-linear function (ReLU) as shown in Equation 8 below.

According to various embodiments, when the defect representation vector is generated, the classification block 210 may generate a characteristic representation vector using the softmax layer 216 . In this case, the classification block 210 may generate a characteristic expression vector by normalizing the defect expression vector in a probability form using a softmax function as shown in Equation 9 below.

According to various embodiments, when the characteristic expression vector is generated, the classification block 210 may diagnose the partial discharge based on the characteristic expression vector. According to an embodiment, the classification block 210 may diagnose the partial discharge based on the characteristic expression vector and the defect reference model. The defect reference model may be data in which a defect type and a partial discharge model are predefined. For example, the classification block may analyze the partial discharge signal detected by the sensing unit 130 using the characteristic expression vector, and compare the partial discharge signal with a defect reference model to diagnose the partial discharge.

Additionally, the classification block 210 uses a cross-entropy loss function as in Equation 10 and Equation 11 below to optimize the learning parameters, a parameter that minimizes the cost function. can also be derived.

According to various embodiments of the present disclosure, the electronic device includes a memory storing a learning model and an analyzer configured to diagnose partial discharge using the learning model, wherein the analyzer includes a predefined partial discharge signal pattern (Phase-resolved PD). ) to generate embedding data, extract features from the embedding data based on a self-focusing technique, and diagnose partial discharge based on the extraction result.

According to various embodiments, the embedding data may include a query vector, a key vector, and a value vector. For example, the predefined partial discharge signal pattern may include data obtained through a simulation failure test.

According to various embodiments of the present disclosure, the analysis unit may be configured to calculate an attention value from the embedding data and apply a residual connection to the attention value.

According to various embodiments of the present disclosure, the analyzer may be configured to linearly transform the attention value to which the residual connection is applied using a nonlinear function, and to apply the residual connection to the linearly transformed attention value.

According to various embodiments of the present disclosure, the analysis unit may be configured to generate a characteristic expression vector for the partial discharge diagnosis, based on the extraction result.

According to various embodiments, the analysis unit may be configured to adjust the size of the extracted feature data using a pooling technique.

According to various embodiments, the analyzer may be configured to generate the scaled data as a defect expression vector by using a nonlinear function.

According to various embodiments, the analyzer may be configured to normalize the defect expression vector to a probability form using a softmax function.

According to various embodiments, the analyzer may be configured to optimize learning parameters based on the loss function and the feature expression vector.

3 is a flowchart illustrating a partial discharge diagnosis operation of the electronic device 100 according to various embodiments of the present disclosure. In the following embodiment, each operation may be performed sequentially, but is not necessarily performed sequentially. In addition, the following operations may be performed by the processor 110 of the electronic device 100 or implemented as instructions executable by the processor 110 .

Referring to FIG. 3 , the electronic device 100 according to various embodiments may acquire a partial discharge signal in operation S310 . According to an embodiment, the partial discharge signal may be data obtained from a simulated failure test.

According to various embodiments, in operation S320 , the electronic device 100 may generate embedding data based on the partial discharge signal. The embedding data may be a vector having N real values, and may include a query vector, a key vector, and a value vector. According to an embodiment, the electronic device 100 may generate a key vector, a value vector, and a query vector based on the partial discharge signal. For example, such embedding data may be generated by the self-concentrating block 200 .

According to various embodiments of the present disclosure, in operation S330 , the electronic device 100 may extract a feature of the embedding data based on the self-concentration technique. The feature of the embedding data may be calculated by an attention value, as will be described later with reference to FIG. 4 . According to an embodiment, the electronic device 100 may extract a feature of the embedding data based on the aforementioned <Equation 1> to <Equation 4>. For example, the features of such embedding data may be extracted by the self-concentration block 200 .

According to various embodiments of the present disclosure, in operation S340 , the electronic device 100 may diagnose partial discharge based on the characteristics of the embedding data. According to an embodiment, the electronic device 100 may diagnose the partial discharge using the partial discharge signal detected by the sensing unit 130 and the features extracted from the embedding data. For example, diagnosing the partial discharge may include determining whether the partial discharge has occurred. For example, the electronic device 100 may determine whether partial discharge has occurred or the degree of occurrence of partial discharge based on a feature extracted from the embedding data. As another example, diagnosing the partial discharge may include determining the type of partial discharge. For example, the electronic device 100 may generate at least one of Particle, Floating, Corona, Viod, and Noise based on the feature extracted from the embedding data. can be diagnosed. For example, such a diagnosis may be performed by the classification block 210 .

4 is a flowchart illustrating an operation of extracting a feature of embedding data in the electronic device 100 according to various embodiments of the present disclosure. The operations of FIG. 4 described below may represent various embodiments of the operation S330 of FIG. 3 . In addition, in the following embodiments, each operation is not necessarily performed sequentially, and at least one operation among the disclosed operations may be omitted or another operation may be added.

Referring to FIG. 4 , the electronic device 100 according to various embodiments may calculate an attention value for embedding data in operation S410 . As described above, the embedding data may include a query vector, a key vector, and a value vector. For example, the electronic device 100 may calculate an attention value indicating a correlation between a query vector and a key vector.

According to an embodiment, the electronic device 100 uses the self-concentration block 200 to scale a similarity obtained by performing a dot-product of a query vector and a key vector, and uses this as a softmax Attention value can be calculated by applying to the (softmax) function, normalizing it, and then multiplying it with a value vector. Such an attention value may be calculated through <Equation 1> and <Equation 2> as described above with reference to FIGS. 2A and 2B .

According to various embodiments, the electronic device 100 may apply the residual connection to the calculated attention value in operation S420 . The residual connection is to prevent the learning data learned in the self-concentration sub-block 202 from being lost in the data processing process, and the electronic device 100 is Similarly, the residual concatenation of adding input data (eg, embedding data) to the output data of the self-concentrated sub-block 202 may be applied.

According to various embodiments of the present disclosure, in operation S430 , the electronic device 100 may linearly transform the attention value to which the residual connection is applied and output it. According to an embodiment, the electronic device 100 may linearly transform the attention value to which the residual connection is applied using a non-linear function (ReLU) and output it as a classification block. In this case, the electronic device 100 may apply the residual connection to the linearly transformed attention value. For example, as described above with reference to FIGS. 2A, 2B and <Equation 4>, the residual linkage may be applied by adding the linearly transformed attention value and the nonlinearly transformed residual linkage applied attention value.

5 is a flowchart illustrating an operation of diagnosing partial discharge in the electronic device 100 according to various embodiments of the present disclosure. The operations of FIG. 5 described below may represent various embodiments of operations S330 to S340 of FIG. 3 or operation S430 of FIG. 4 . In addition, in the following embodiments, each operation is not necessarily performed sequentially, and at least one operation among the disclosed operations may be omitted or another operation may be added.

Referring to FIG. 5 , the electronic device 100 according to various embodiments may adjust the size of input data in operation S510 . The input data may be data outputted after being linearly transformed in operation S430 of FIG. 4 . According to an embodiment, as described above with reference to FIG. 2C and <Equation 7>, the electronic device 100 divides the input data into filters of a predetermined size and selects the largest value in the filter (maximum pooling ( max pooling), the size of the input data can be reduced. As described above, the maximum polling technique is an example for reducing the size of input data, and the present disclosure is not limited thereto. For example, various types of pooling techniques capable of preventing overfitting of input data and reducing the number of parameters may be used to adjust the size of input data.

According to various embodiments, in operation S520 , the electronic device 100 may generate a characteristic representation vector based on the size-adjusted input data. According to an embodiment, the electronic device 100 may generate a fault representation vector using the nonlinear function of the scaled input data, and apply it to the softmax function to generate a characteristic representation vector. there is. For example, the defect expression vector may be generated based on the aforementioned <Equation 8>, and the characteristic expression vector may be generated based on the aforementioned <Equation 9>.

According to various embodiments of the present disclosure, in operation S530 , the electronic device 100 may diagnose partial discharge based on the characteristic vector. According to an embodiment, the electronic device 100 may diagnose the partial discharge based on the characteristic expression vector and the defect reference model. The defect reference model may be data in which a defect type and a partial discharge model are predefined.

Additionally, the electronic device 100 may optimize learning parameters. For example, as described above with reference to FIG. 2C , <Equation 10> and <Equation 11>, the electronic device 100 calculates a parameter for minimizing the cost function using a cross-entropy loss function. can also be derived.

6A and 6B are diagrams for explaining the performance of partial discharge diagnosis according to the present disclosure.

The partial discharge diagnosis method according to the present disclosure may use a partial discharge signal as an input and extract and learn a correlation between an input and an output. It was confirmed through experiments that the diagnostic performance and algorithm complexity of this partial discharge diagnostic method were improved compared to the existing RNN-based partial discharge diagnostic method.

Specifically, referring to FIG. 6A , in order to test the diagnostic performance, five types of simulated failure data obtained through a simulation failure experiment as indicated by reference numeral 610 were utilized, and in order to diagnose partial discharge according to the present disclosure, reference numerals The same model parameters as 620 were used.

Referring to FIG. 6B , as a result of the experiment, as indicated by reference numeral 630 , it was confirmed that the performance of partial discharge diagnosis by the method according to the present disclosure was superior to that of partial discharge diagnosis using the existing diagnosis method.

Also, as indicated by reference numeral 640, it can be seen that the diagnostic method according to the present disclosure can use a smaller number of parameters compared to the existing diagnostic method, and requires less training and learning time.

According to various embodiments, a method for diagnosing partial discharge in an electronic device includes an operation of generating embedding data based on a predefined partial discharge signal pattern (Phase-resolved PD), and an operation of generating embedding data from the embedding data based on a self-focusing technique It may include an operation of extracting a feature and an operation of diagnosing the partial discharge based on the extraction result.

According to various embodiments, the embedding data may include a query vector, a key vector, and a value vector. According to an embodiment, the generating of the embedding data includes generating the key vector and the value vector based on the partial discharge signal and based on a predefined partial discharge signal pattern (Phase-resolved PD). It can include the operation of creating a query vector.

According to various embodiments, the predefined partial discharge signal pattern may include data obtained through a simulation failure test.

According to various embodiments, the operation of extracting the feature from the embedding data may include calculating an attention value from the embedding data and applying a residual connection to the attention value.

According to various embodiments, the operation of extracting the feature from the embedding data includes an operation of linearly transforming an attention value to which the residual connection is applied using a nonlinear function and an operation of applying the residual connection to the linearly transformed attention value. may include

According to various embodiments, the operation of diagnosing the partial discharge may include the operation of generating a characteristic expression vector for diagnosing the partial discharge, based on the extraction result.

According to various embodiments, the operation of generating the feature expression vector may include the operation of adjusting the size of data obtained through the operation of extracting using a pooling technique.

According to various embodiments, the generating of the feature expression vector may include generating the scaled data as a defect expression vector using a nonlinear function.

According to various embodiments, the operation of generating the feature expression vector may include the operation of normalizing the defect expression vector into a probability form using a softmax function.

According to various embodiments, the operation of diagnosing the partial discharge may include an operation of optimizing learning parameters based on a loss function and the characteristic expression vector.

The partial discharge diagnosis method according to embodiments of the present disclosure may be implemented with instructions that are stored in a computer-readable storage medium and executed by a processor.

A storage medium, whether directly and/or indirectly, in a raw, formatted, organized or any other accessible state, may include a relational database, a non-relational database, an in-memory database, Alternatively, it may include a database, including a distributed one, such as any other suitable database capable of storing data and allowing access to such data through a storage controller. In addition, the storage medium includes a primary storage device (storage), a secondary storage device, a tertiary storage device, an offline storage device, a volatile storage device, a non-volatile storage device, a semiconductor storage device, a magnetic storage device, an optical storage device, a flash. It may include any type of storage device, such as a storage device, a hard disk drive storage device, a floppy disk drive, magnetic tape, or other suitable data storage medium.

Although the present disclosure has been described with reference to the embodiment shown in the drawings, this is merely exemplary, and those of ordinary skill in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Accordingly, the true technical protection scope of the present disclosure will have to be determined by the technical spirit of the appended claims.

Claims (20)

A method for diagnosing partial discharge in an electronic device, comprising:
generating embedding data based on a predefined partial discharge signal pattern (Phase-resolved PD);
extracting a feature from the embedding data based on a self-focusing technique; and
and diagnosing the partial discharge based on the extraction result.
The method of claim 1,
The embedding data includes a query vector, a key vector, and a value vector.
The method of claim 1,
The predefined partial discharge signal pattern includes data obtained through a simulation failure test.
The method of claim 1,
The operation of extracting features from the embedding data includes:
calculating an attention value from the embedding data; and
and applying residual linkage to the attention value.
5. The method of claim 4,
The operation of extracting features from the embedding data includes:
linearly transforming the attention value to which the residual connection is applied by using a nonlinear function; and
and applying a residual linkage to the linearly transformed attention value.
The method of claim 1,
The operation of diagnosing the partial discharge includes:
and generating a characteristic expression vector for the partial discharge diagnosis based on the extraction result.
The method of claim 1,
The operation of diagnosing the partial discharge includes:
and adjusting the size of data obtained through the extracting using a pooling technique.
8. The method of claim 7,
The operation of generating the feature expression vector comprises:
and generating the scaled data as a defect expression vector by using a non-linear function.
9. The method of claim 8,
The operation of generating the feature expression vector comprises:
and normalizing the defect expression vector to a probability form using a softmax function.
7. The method of claim 6,
The operation of diagnosing the partial discharge includes:
and optimizing learning parameters based on a loss function and the feature expression vector.
In an electronic device,
memory to store the training model; and
an analysis unit configured to diagnose partial discharge using the learning model;
The analysis unit,
Generating embedding data based on a predefined partial discharge signal pattern (Phase-resolved PD),
extracting features from the embedding data based on a self-focusing technique,
an electronic device configured to diagnose partial discharge based on the extraction result.
12. The method of claim 11,
The analysis unit,
The electronic device is configured to generate the embedding data including a query vector, a key vector, and a value vector based on the predefined partial discharge signal pattern.
12. The method of claim 11,
The predefined partial discharge signal pattern includes data obtained through a simulation failure test.
12. The method of claim 11,
The analysis unit,
calculating an attention value from the embedding data,
an electronic device configured to apply residual concatenation to the attention value.
15. The method of claim 14,
The analysis unit,
Linear transformation of the attention value to which the residual connection is applied using a non-linear function,
an electronic device configured to apply a residual linkage to the linearly transformed attention value.
12. The method of claim 11,
The analysis unit,
The electronic device is configured to generate a characteristic expression vector for the partial discharge diagnosis based on the extraction result.
12. The method of claim 11,
The analysis unit,
An electronic device configured to adjust the size of the extracted feature data using a pooling technique.
18. The method of claim 17,
The analysis unit,
An electronic device configured to generate the scaled data as a defect expression vector by using a nonlinear function.
19. The method of claim 18,
The analysis unit,
An electronic device configured to normalize the defect expression vector to a probability form using a softmax function.
17. The method of claim 16,
The analysis unit,
An electronic device configured to optimize learning parameters based on a loss function and the feature expression vector.

Images